Where a model was made and where it runs are separate questions
A request for "European AI" mixes two questions: who built the model, and where your prompts are processed. They have different answers and different laws behind them.
Do we need a European-made AI model for GDPR, or is European hosting enough? The question usually comes from a privacy team or a customer's security questionnaire. Someone has heard that only a European model is safe, and the engineering team has to answer.
An open-weight model is one whose weights, the numbers that make up the trained model, are published for download. That changes the picture. You can take the same file and run it on GPUs anywhere. So the place it was trained says nothing about where your next prompt goes.
Lyceum is a sovereign European AI inference provider for open-source models. We build and run our platform in Europe, and our GPUs run in European data centres. Our dedicated endpoints serve that model for you alone, and the data centre is agreed for your workload.
The table below splits the request into the two checks your legal team will run.
| Question | Which law asks it | Who has to answer |
|---|---|---|
| Where are prompts processed, and by whom? | GDPR, Articles 28 and 44 | You, as controller, and your inference provider, as processor |
| What does the model's documentation say about training and testing? | EU AI Act, Article 53 | The model's maker, wherever it is based |
What GDPR checks: who processes your prompts, and where
GDPR follows the personal data inside your prompts. Two roles matter here. The controller is the company that decides why the data is processed, usually you. The processor is the company that handles the data on your behalf, such as an inference provider.
Article 28 says a controller may use only processors that give sufficient guarantees through technical and organisational measures. These are the security and access controls a provider puts in place. Article 44 adds a rule for location. A transfer means, roughly, sending personal data to a recipient in another country. A transfer to a third country has to meet the conditions of GDPR Chapter V, the chapter on transfers. A third country is one outside the European Economic Area (EEA): the EU plus Iceland, Liechtenstein and Norway.
Neither article mentions where a model was trained. A model made in Europe but served from outside the EU still raises the transfer question. A model made elsewhere but served in a European data centre may not raise the transfer question at all. That holds as long as no one sends the prompts on to a company outside the EEA.
So the useful questions for a provider are concrete. In which country does inference run? Which company processes the prompts? Is anything stored after the response?
Our own answers sit on our legal page. A DPA, or data processing agreement, is the contract that sets how a processor handles your data. Under our DPA, you are the controller and we are the processor, and the DPA is Annex 1 of our terms. We don't keep your prompts or outputs, and we never use them for training. A subprocessor is another company a processor uses for part of the service. You can object to a new subprocessor within 10 business days of our notice.
What the AI Act asks of the model's maker
The EU AI Act looks at the model itself. The Act calls the company that makes a general-purpose AI model its "provider". Here we say model maker, to keep it apart from the inference provider that runs the model for you. The Act covers model makers who place their models on the EU market, meaning make them available in the EU. It applies whether they are established in the EU or in a third country.
Those model makers carry their own duties. Article 53 requires them to draw up and keep up to date the model's technical documentation, including its training and testing process.
Two things follow for you. The AI Act still applies to a model maker based outside Europe. And being made in Europe does not replace the documentation the Act asks for. In both cases, the question to ask is what the maker publishes about the model.
This is a reading of the two texts, not legal advice. Your own counsel decides how they apply to your use case.
Sizing a downloaded open-weight model for one GPU
Once origin is out of the way, the practical question is whether the model fits the hardware you want to reserve. Take a hypothetical model from a European lab with 24 billion parameters, downloaded from Hugging Face.
A model's parameters are its weights: the learned numbers that make up the model, counted in billions. Each weight is stored at a set precision, meaning the number of bits per weight. GPU memory, or VRAM, has to hold the weights. It also needs room for the prompts the model is processing at that moment. The weights alone follow one formula:
weight memory (GB) = parameters (billions) x bytes per weight
bytes per weight: 16-bit = 2, 8-bit = 1, 4-bit = 0.5For our hypothetical 24-billion-parameter model on one NVIDIA H100, a data-centre GPU with 80 GB of memory:
| Precision | Weight memory | Left on an 80 GB GPU |
|---|---|---|
| 16-bit | 24 x 2 = 48 GB | 32 GB |
| 8-bit | 24 x 1 = 24 GB | 56 GB |
| 4-bit | 24 x 0.5 = 12 GB | 68 GB |
Calculated example for a hypothetical model; weights only.
The same model fits one GPU at every precision. Nothing in this calculation depends on where the model was made. What is left to decide is where it runs and under which contract.
The example is optimistic, because it counts the weights only. It does not show two things. Long prompts and many requests at once need more of the remaining memory. Lower precision can also change the model's answers, so test it on your own tasks first. Our sizing tool shows the memory a Hugging Face model needs and how many requests it serves.
Running any Hugging Face model in a European data centre
A dedicated endpoint is a model running on GPUs reserved for you. A replica is one running copy of the model on its own reserved GPUs. Ours run the model on replicas behind an OpenAI-compatible Chat Completions endpoint. You reference the model by its Hugging Face model ID, so a European-made model and any other open model take the same path.
The command below deploys a model from our CLI. The model ID is a placeholder for the one you choose.
lyceum infer deploy your-org/your-model -g gpu.h100 --min-replicas 1 --max-replicas 3For a gated model, one that needs approval to download, you add a Hugging Face token with the -t flag. The details are in our dedicated inference docs.
Where a dedicated endpoint runs is agreed per contract, for your workload. You can tell us if you need a specific EU location. Our serverless models, the shared models billed per token, run in Paris and Finland.
On a dedicated endpoint, billing follows the time the replicas run, not each request. While traffic flows, the endpoint keeps at least the number of replicas set with --min-replicas. After 1 hour with no requests, it scales down to zero and is paused. The next request resumes it at that minimum, so the first answer after a pause takes longer.
A checklist for your legal review
Bring these questions to your legal team, and ask every provider the same ones, us included.
- In which country are prompts processed for the product you will use?
- Which company processes them, and under which DPA?
- How are you told about a new subprocessor, and how can you object?
- Are prompts or outputs stored, logged or used for training?
- Separately: what technical documentation does the model's maker publish?
Questions 1 to 4 are about hosting. Only question 5 is about the model, and it applies to models from Europe and elsewhere alike.
Say the first four answers point to processing in a European data centre, under a DPA you have read. Then the model's country of origin stops being the deciding factor. Enter your model in our sizing tool to see the memory it needs. Then book a call about a dedicated endpoint in the EU location you need.
Frequently asked questions
Which rules look at my prompts, and which look at the model's maker?
GDPR looks at your prompts: who processes the personal data in them and whether it goes to a country outside the EU. The AI Act's documentation duty for general-purpose models looks at the model's maker.
How quickly does Lyceum add newly released open-weight models, and can customers request a specific model?
As fast as we can. Our team starts on new open-weight models as soon as the weights are released, and we prioritise based on model quality and customer demand. If you need a specific model, tell us and we'll look at adding it.